⚡ AI 情报站

We're extending access to Claude Fable 5 on all paid plans through July 12.

GPT-5.6 Sol, along with Terra and Luna, will launch publicly this Thursday.

We’re expanding preview access globally now.

Claude Cowork is coming to mobile and web.

Hand Claude a task at your desk and pick up the finished work from your phone. Close the laptop and Claude keeps going.

Beta is rolling out over the next several weeks starting with the Max plan, with more plans to follow.

GPT-5.6 sol launches thursday!

happy building

I can finally talk about 5.6.

I’ve been testing it for months and, without exaggeration, it’s the best model I’ve ever used.
Fast, smart, genuinely creative, and you guessed it, they finally fixed front-end design.

I haven’t needed to check the code I’ve written in two months.
We’ve reached the point where, if you can imagine something, you should just build it.

As before, you can use up to 50% of your weekly usage limit on Claude Fable 5. After that, you can keep using Fable 5 with usage credits, or switch to another model to keep working within your remaining limits.

Read more:
support.claude.com/en/articl…

Introducing Muse Image and Muse Video, the first media generation models developed by Meta Superintelligence Labs.

Muse Image is our most advanced image generation model yet. It follows instructions faithfully, edits with precision, composes from multiple references, and draws on Instagram for social context. It also brings agentic tool use capabilities to image generation and integrates with Muse Spark.

You can try Muse Image in the Meta AI app and web, as well as in Instagram Stories and WhatsApp – starting in limited countries with more locations on the way.

Today we’re also previewing Muse Video, which is built upon the same pretraining base as Muse Image to deliver exceptional visual fidelity with native audio support.

Learn more about both models:
go.meta.me/7f4427

🇪🇺 The EU is now for the 6th time trying to force Chat Control through which lets them scan ALL your private messages, photos and emails without a warrant

Implictly showing the EU is not democratic and not about what the people of Europe want, because once a law is rejected, you just re-submit it until nobody is watching and it's passed

November 2023: ❌ Chat Control is rejected
June 2024: ❌ Chat Control is rejected
October 2025: ❌ Chat Control is rejected
November 2025: ❌ Chat Control is rejected
March 2026: ❌ Chat Control is rejected
July 2026: 📝 Chat Control is back

Even the EU's own lawyers stated Chat Control is unconstitutional: "generalised message scanning is incompatible with Article 7 of the EU Charter"

You have to wonder why the EU is so adament about reading your private chats, right?
So
@marckohlbrugge
told me to finally make a Nomads iOS app today

But I didn't want to do it locally because I only Claude Code on VPS now, so I asked it how to do it and it suggested to rent a
@MacinCloud
(not affiliated)

So I did, then handed the login to Claude on my VPS and now it's coding an iOS app via SSH with Xcode on a Mac Mini somewhere in California! 🤯

One big reason to only code on the server for me is I wanna be able to work on my iPhone a bit or at least track progress when I'm away from home and it's annoying lugging my MacBook Pro 16" around everywere, so I don't anymore

My work now is about 50-50 on my iPhone or MacBook Pro, both via Termius SSH to VPS! It's lovely! 🥰
I'd just make a second page for me to test it before I deploy it then
Because that means you still have to deploy to a web server

Why not just work on the web server?

Crazy concept :D

I had early access to GPT-5.6 Sol.

It’s an amazing model, but for almost every task I tested, Fable was quite a bit better, and more agentic to boot (one Fable turn does the same thing many 5.6 turns do).

Full review coming on launch day!
Don’t get me wrong, this model is a huge jump over 5.5.

Fable just has the feel of a much bigger model that I can trust a good bit more!

I expect OpenAI will have an answer to Fable soon.

Your Favorite Model, Now In Your Favorite Agent

Kling MCP & CLI is officially live!
Now, you can call Kling directly within your favorite AI Agents. Your go-to Skills are all here—no switching tools frequently, no move assets around manually. With just one sentence, you can hand your creative tasks over to your “AI assistant.”
Strong realism, high-fidelity details, and cinematic aesthetics—e-commerce sellers, small business, and content creators can all produce at scale with low cost, double their efficiency, and keep inspiration flowing.
Get started with Kling MCP and build your unique AI creative workflow.

ONE MORE WEEK TO ESCAPE THE UNDERCLASS

I had Fable build another thing I always wanted, a full procedural fantasy kingdom generator with economics, trade routes, population growth, wars, lineages, and occasional dragons. First, I worked with it on a plan, then it made it.

You can play it here:
annals-kingdom.netlify.app/

🤯 And it worked!

It made a basic iOS app in full Swift then made a screenshot and sent me a link to check it!

It made it on my Hetzner VPS but then SSH'd into a
@macincloud
Mac Mini in a datacenter somewhere to then build an app with Xcode on there

Very basic but a great start

I was an early tester of GPT-5.6 Sol. I was asked to not share demos until after launch but it is a very good model.

It is of similar ability, but quite different feel, than Fable. Fable wants to go off and do work on its own pace, Sol is faster but works with you in steps more.
I found myself switching between Fable and Sol depending on task. Sol for back-and-forth tasks, especially when I had not yet figured out what I needed exactly, Fable for very long tasks where I could define what I wanted, and Sol Pro for really hard problems.

Sol is rising. It’s a good model.

Never ever buy a Daikin AC!

They limit how many times you can change them to just 150 times per day!

We have 7 ACs in the house so that's a limit of 21 instructions per day 😂 completely arbitrary btw, makes no sense, just classic Japanese software design

(I didn't buy it, it came with the house, and I've actively been ripping them out and replacing them with Mitsubishi)

Other good brands for ACs are LG, Hisense and Haier, can recommend!

we're eating good tonight
Going to bed now, but I will post more tomorrow. Ultra is low-key insane.
Another thing I want to add is how incredible
@OpenAI
employees are. They were super helpful as we navigated and tried to understand what was going on during the government halt. And they really care about our feedback. It’s such a wonderful group of people.

Announcing Robostral Navigate, our first model for embodied navigation: an 8B robotics navigation model that guides robots to autonomously perform tasks specified with natural language. Single RGB camera. State-of-the-art on R2R-CE.

Alongside the release of Muse Image, we’re sharing an early preview of Muse Video. It offers competitive performance in prompt adherence, visual fidelity, and temporal consistency.

We’re investing in areas with current performance gaps, such as audio-video synchronization and physically accurate fast motion.

And it's made profile pages now, and because Claude Code runs on same server as the site itself it can instantly build an /api route that the iOS app then uses to get the data to put everything in native Swift elements

I didn't touch anything, I just prompt, and I also can't even login to check on it because it runs headless on the Mac Mini in the cloud! Very cool

It's so easy to make an iOS app with AI!!! I'm finally not iOS-retarded anymore 😊😊😊

Very excited to help chart the future of Git (and SCM generally) for the agentic future with Taylor!

don’t forget to overthink the best case scenario too : )

My big takeaway is that both Sol & Fable represent jumps over previous models and have opened a large gap with the next-best AIs. People will have preferences for one or the other, but if you doing any work where better intelligence matters, those two models are your only choices

Hit my
@Cloudflare
Email Sending quota today 😭

Can't send any emails on any of my sites anymore

Chat and Cowork are moving into one home on web and desktop, with one place for your projects and artifacts across both.

Handing Claude a task starts the same way a conversation does.

Today we are rolling out some big updates to Managed Agents in the Gemini API:

- support for background tasks
- remote MCP & function calling
- network credential refresh

and you can now get started with Managed agents in the API via the free tier!

We're extending doubled Cowork usage limits through August 5, so you can delegate bigger work to Claude.

Read more:
claude.com/blog/cowork-web-m…

Another little thing that nothing but Fable could have helped me make today:

You can now drag and drop files from your own computer into Windows 3.11 on
pieter.com


The file is saved inside C:\DOCS and if you want to share it with other people you can put it in C:\SHARED

It all works with a web filesystem called WebFS and it's very very cool, kind of becoming like a shared library/school computer

Codex for creating the first video of a black hole:

Scheduled tasks now run even with your computer off. Set client prep for 6am, and Claude works through the threads and transcripts, builds the brief, and leaves the follow-up drafted but unsent.

When a decision needs your judgment, the question comes to your phone.

Anthropic's new J-Space paper is fascinating.

It describes an internal workspace where Claude keeps concepts in mind before they appear in its response.

Inspired by the paper, I built a skill that lets Claude (and others) reveal their J-Space.
Here's what it looks like.🧵

Vera CPUs are custom-built for agentic runtime.
We've been working together with NVIDIA on running our sandbox infrastructure that powers Perplexity Computer on Vera CPUs, with significant improvements. We will be sharing more on this soon!

Calling all solopreneurs and small business owners! 📣

Join our next live demo to learn how Gemini can help power your workflows and your growth. A Google Product Manager will walk you through how to unlock personalized insights, automate repetitive tasks, and brainstorm new content for your business.

👉Join the Discord to watch live:
discord.gg/gemini

📅 Tomorrow (Wednesday, July 8) at 11:30am PT

Muse Image works as an agent rather than a direct prompt-to-image model: it invokes tools, self-refines, improves with scaled test-time compute, and pairs with Muse Spark for collaborative media generation.

🧵👇

GLM-5.2 on Ollama's cloud just got more capacity in US & Europe!

Ollama's cloud for GLM 5.2 consistently delivers between 80 to 120 output tokens per second, even during peak hours, compared to 30 to 40 tok/s on other providers.

Use GLM-5.2 with

Claude Code:
ollama launch claude

VS Code
ollama launch vscode

Codex
ollama launch codex

Hermes
ollama launch hermes

it_happening.gif

Agreed.

Nothing else (sorry, local AI folks) compares to Fable and Sol.

The overwrought language of Fable is an ongoing problem in software and design projects where the AI is otherwise excellent. It just creeps into little bits of text and toasts and menus, even if you try to instruct otherwise. And once it gets in it is hard to purge completely.

Short Video Overviews are officially 💯 rolled out on mobile and web across all users in English!

As always, your opinion means everything to us. Share your favorite creations below and tell us what features we need to add next! ❤️

MAI-1 has no independent benchmarks yet, but the ones they released suggest it is worse than Sonnet 4.6.

Not sure it is going to be great as a Copilot at Excel and Outlook, especially as there are already quite good Claude/OpenAI plugins for office using good AI. We will see.

Crazy to see what happened to
@GroqInc
after my viral post.

Jonathan spent years trying to explain why fast inference matters. One post of mine finally made people feel it.

I’ve done the same for Rork, OpenPipe and others.

This is my main value-add as an investor!

Anthropic: "We're extending access to Claude Fable 5 on all paid plans through July 12."

Me: 💀

Claude Cowork just escaped the laptop.

Start work at your desk.
Check it on your phone.
Claude keeps going in the background.

This is agentic work.

Just found out about this from this little popup message in the Claude chat web UI

Today, we're launching Runway Dev, a new AI media platform for professional developers and enterprise teams.

From search and advertising to ecommerce and gaming, developers are using Runway to cut costs, iterate faster and drive better consumer engagement. Runway Dev gives teams building generative media experiences a single platform from one vendor, removing the need to stitch together a dozen disparate APIs on their own.

Learn more at the link below.

If you care about this follow
@MartinSonneborn
who's fighting for your privacy and freedom!

Prompt: ANNALS — a living kingdom in a single file

Working title behavior: the app titles itself per seed — "The Annals of Vaelmere", "The Annals of Osterholt" — because the central conceit is that you are watching a history write itself. The kingdom is the simulation; the Chronicle is the soul.

1. Vision

A procedurally generated fantasy kingdom rendered in real-time 3D that you can watch for five minutes or five hours. Zoom from a satellite view of the whole realm down to a market stall with laundry on the line. Time runs; harvests come in; caravans crawl the roads; a queen dies without a clear heir and two great houses raise banners; plague follows the trade routes; a dragon notices the treasury has gotten large. Everything that happens is written into a scrolling Chronicle in the voice of a medieval annalist, and a cinematic auto-director flies the camera to the story as it unfolds.

Four pillars, in priority order:

Legible emergence. Systems couple visibly: drought → failed harvest → hunger → unrest → rebellion. A viewer should be able to reconstruct cause from effect just by watching.

Zoom is the reward. Every altitude has something worth seeing — political geography at 3 km, trade arteries at 800 m, street life at 40 m.
The Chronicle is the product. Named people, named places, consequences that persist (ruins, grudges, epithets). The sim generates a story you could retell.
Watchable by default. With zero input it should behave like a nature documentary about a kingdom.
2. Hard constraints

Single HTML file, vanilla JS + Three.js from CDN (write against the r128 API surface so the same file runs as a Claude artifact during development and unchanged on Netlify). No build step, no framework, no backend.

Custom camera controller (no OrbitControls dependency — we need a bespoke rig anyway, see §5).

No localStorage / sessionStorage. Persistence is the seed: URL hash
#s
=1234567 fully determines the generated world. "Copy link" button shares it. Optional "Export chronicle" downloads a .txt.
Deterministic worldgen via seeded sfc32 RNG with separate streams for generation and history, so intervening in history never changes the map.
Performance floor: 60 fps on a mid-tier laptop at every altitude; <250 draw calls via instancing and merged geometry; graceful degradation, never a crash.
Everything procedural — no external textures or models. Vertex colors, tiny generated canvas atlases (windows, sigils), and shader work carry the visuals.
3. World generation pipeline
Runs once per seed, in order, with a brief "forging the realm" progress line:

Terrain. 6×6 km world. fBm heightmap with domain warping; one mountain spine, rolling lowlands, ~40% of seeds get a coastline on one edge.

Hydrology. Flow accumulation carves one major river (always) plus tributaries; lakes in basins; river mouths become harbor candidates.

Biomes. Moisture × elevation → meadow, farmland-suitable plain, deciduous forest, pine upland, rocky peak, marsh. Drives tree instancing, ground vertex color, and resource nodes (farmland, forest, ore vein in rock, fishing water).

Settlement siting. Score cells by water access, flat buildable area, resource adjacency, harbor bonus. Place 1 capital (pop ~2,500), 2–3 towns (400–1,200), 3–5 villages (60–300). Capital gets a keep on defensible high ground and a walled old core.
King's roads. A* over a cost field (slope² penalty, water crossing penalty that spawns bridges, forest penalty) connecting all settlements into a network. Roads are the circulatory system — trade, armies, plague, and bandits all use them.
Streets & lots. Inside settlements, organic street growth seeded from the road entry points and the market square; lots subdivide along streets; districts assigned (market, artisan, residential, temple, keep). Villages are a crossroads, a well, and a cluster.
Buildings. ~2,500–4,500 total from 12 archetypes × 3 tiers (hovel→timber→stone): house, longhouse, shop, smithy, mill (animated wheel on watercourses), granary, tavern, temple, warehouse, tower, wall segment/gate, keep. Procedural assembly: footprint, floors, roof style, chimneys, timber-frame overlay at tier 2+, hanging signs on shops.
Names & heraldry. One coherent phonology: places from syllable banks with earthy suffixes (-mere, -ford, -holt, -caster, -wyck); people from gendered given-name banks; great houses take a place or beast name. Procedural sigils = field color pair + one charge from a bank of 16 (stag, tower, wave, oak, wolf, comet, key, sun…), rendered to a small canvas atlas and used on keep banners, army flags, and UI chips.
The cast. Crown dynasty (monarch, consort, children) + 5 great houses, each holding a region and seat. ~100–140 living notables at any time (see §7).
4. Rendering, LOD & visual style
Aesthetic: illuminated-manuscript storybook. Low-poly forms with painterly vertex-color gradients — not flat-shaded minimalism, not photorealism. The world should look like the marginalia of a chronicle came to life. One directional sun/moon light with a tight camera-tracking shadow frustum (shadows fade out above 1,500 m altitude), height + distance fog for depth, custom gradient sky dome with sun, moon, stars, and the occasional comet.
Water: gerstner-lite vertex displacement + fresnel tint on rivers, lakes, sea. Rivers visibly swell during floods.

Day/night: full cycle; nights are deep indigo with warm emissive windows (per-instance emissive flag in the bubble band), pooled torch/lamp point lights (≤12), keep braziers.

Seasons (90 days each, 360-day year): global palette lerp + effects. Spring: saturated greens, snowmelt-fat rivers. Summer: warm golds (drought pushes to straw-yellow). Autumn: russet forest ramp. Winter: snow accumulation via normal-up whitening in the ground/roof shaders, snowfall particles, frozen lake tint.

Weather: clear/overcast/rain/storm/snow state machine; rain streak particles, wet-darkened ground tint, storm lightning flashes. Weather is mechanical, not just decorative (fire spread, battle modifiers, flood meter).

Ambient life: chimney smoke curls, birds (cheap boid flock), grazing sheep sprites near farms, banner cloth idle-wave shader, market crowd murmur of villagers in the bubble band.
LOD bands (camera distance to object):
Band Range Buildings Agents Detail Far >2,200 m merged/instanced color-block masses per district armies → banner icon, caravans → moving dot settlement name labels, territory tint Mid 500–2,200 m instanced archetype meshes full low-poly agent meshes instanced trees, roads, bridges Near 120–500 m + chimneys, signs, timber overlays walk-bob animation wells, carts, market stalls Bubble <120 m + per-instance window glow, laundry lines, lamps citizens spawn (≤80), individual heads turn prop shadows, torch lights
Citizens outside the bubble do not exist as entities — they're statistics (§7). The bubble makes the zoom promise real without simulating 8,000 pedestrians.
UI skin: parchment panel (
#E9DCC3
) with iron-gall ink text (
#2B2320
), crown-gold accents (
#C9A227
), house colors from sigils. Chronicle set in a period serif (IM Fell English or similar via Google Fonts, with graceful serif fallback); UI labels in a clean humanist sans. Trigger buttons read like acts, sentence case, active voice: "Unleash plague," "Bless the harvest," "Contest the succession."

5. Camera & interaction
Custom rig with three modes, all sharing one smoothed spring-damper transform:
Free: orbit + pan + scroll zoom along a curved altitude path — high altitude looks steeper (map-like), low altitude tilts toward the horizon (street-like). Zoom range ~5 m to ~4,000 m. Terrain collision keeps the camera out of the ground.
Follow: click any agent (caravan, army, notable, dragon) → low tracking shot with slight lead. Click empty ground to release.
Director: see §13.
Raycast click on anything selectable opens the Inspector (§15). Hotkeys: Space pause, 1–4 speeds, H hide UI, C Watch mode, Esc deselect.
6. Simulation architecture
Clock: SimClock in fractional days. Fixed-step accumulator; 1 tick = 1 sim-day of system logic. Speeds: pause / 0.5 / 2 / 8 / 30 sim-days per real second (at max, a year passes in 12 seconds — a monarch's reign in ~10 minutes).
System order per tick: Weather → Economy → Population → Politics → Military → Threats → Growth → Chronicle flush.
Render decoupling: agents store path + departure/arrival sim-times; the render loop interpolates positions every frame from current sim-time, so motion is smooth at any speed and nothing "teleports."
Stability rule: every positive feedback loop ships with a predator. Growth ↑ → land scarcity and food pressure. Treasury ↑ → dragon and bandit attention. Military strength ↑ → upkeep drain. House power ↑ → crown suspicion events. Target: after 200 unattended sim-years at defaults, population and treasury remain within sane bounds and the realm is still interesting.
Pacing target: a Chronicle-worthy event every 15–45 sim-days at default rates (tunable via the Disaster frequency and House aggression sliders).
7. Population & notables
Two-layer model (the Crusader Kings trick):
Pools: each settlement tracks population as a number with birth/death/migration rates modified by food, plague, war, prosperity. Pools drive building growth, levy size, tax income, and bubble-band citizen density.
Notables: ~100–140 named characters — the dynasty, house heads and their families, plus flavor roles (High Priest, Master of Coin, a famous smith, a bandit king when one arises, a prophet at Myth High). Notables have: age, sex, house, 2 traits (from ~16: bold/craven, just/cruel, shrewd/dull, pious/cynical, fertile, sickly…), relationships (spouse, heirs, grudges), and an agenda line. They age, marry (arranged for alliance value), have children, and die by an age hazard curve plus event hazards (battle, plague, childbirth, the occasional hunting accident, assassination). Deaths and births are Chronicle events; important ones are Director beats.
8. Economy & trade
Legibility over realism. Seven goods: grain, fish, timber, ore, tools, cloth, wine.
Settlements produce by geography (farmland → grain, forest → timber, ore vein → ore, harbor → fish; towns convert ore+timber → tools, capital produces cloth/wine), consume by population, and store in granaries/warehouses.
Local prices from stock vs. need; caravans (mule train meshes, up to ~36 concurrent) spawn to arbitrage price gaps along roads; coastal cogs (up to 8) trade between harbor settlements. Trade volume is visible traffic — the kingdom's health readable at a glance.
Treasury: crown taxes trade and harvests (default 12%); spends on army upkeep, disaster relief, bounties, rebuilding. Treasury funds war; war disrupts trade; disrupted trade starves treasuries.
Hunger cascade: food deficit → granary drawdown → hunger → mortality up, unrest up, migration out. Drought, plague, siege, and raided caravans all feed this one legible pipeline.
9. Dynasty, houses & politics
Succession: gender-blind primogeniture. Heirs need to survive to inherit; a monarch dying with no adult heir triggers a regency (unrest up, house aggression up) or, with rival claimants, a succession crisis.
Legitimacy (0–100) on the crown: raised by coronations, victories, festivals, dragon-slaying; lowered by lost wars, famines under their rule, assassinations, dark omens at Myth High.
House loyalty (0–100 each): drifts with tax rate, shared wars, royal marriages (arranged automatically for alliance value; weddings are festival events), slights and grudges (event-driven), and monarch traits.
Unrest per settlement from hunger, taxes, war, plague, omens → riots (small fires, trade pause) → revolt (settlement flips banner) at extremes.
Civil war trigger: loyalty <30 on a house with military strength + a spark (succession crisis, famine year, tax hike, "Contest the succession" button). Houses declare for sides based on loyalty, marriages, and grudges. Territory overlay animates as banners change.
10. War
Armies: levies raised from settlement pools (strength = f(population, prosperity, walls)); an army is one entity rendered as up to ~300 instanced soldiers with banner (banner-blob icon beyond 800 m). Armies march along roads at ~15 km/sim-day, eat supply, and are gorgeous to follow in Follow cam.
Battles: armies within contact on the same route → 2-sim-day abstract resolution in three phases (skirmish/clash/rout) with visible melee shuffle, arrow arcs, dust, and casualties trickling both pools. Odds from strength × morale × terrain × weather × commander valor. Loser routs toward home; named commanders can die or win epithets ("Aldric the Ironhand").

Sieges: army adjacent to a walled settlement pitches a visible camp (tents, campfires, a trebuchet at scale). Countdown = f(walls, granary, season); sorties and assaults roll on commander boldness. A fallen settlement is sacked (prosperity hit, fire risk, chronicle lament) or occupied (banner flips).

War score from battles, sieges, and months held; at ±100 or exhaustion → peace: new dynasty, exiled house (their sigil struck through in the Houses panel), ceded territory, or a hollow status quo. Ruins and burned districts persist and rebuild over years. Battlefields get a small cairn marker with a tooltip naming the battle. Scars are content.

11. Threats & disasters

Each threat is a coupling into existing systems, not a standalone effect:

Plague: SIR-lite per settlement; imported by caravan arrivals from infected origins (the trade network is the infection network). Virulence and lethality sliders. Visible: pyre smoke columns, district desaturation, tolling pace of the Chronicle. Settlements above 15% infected refuse caravans — emergent quarantine that visibly starves trade.

Great fire: ignition (lightning, riot, sack, drunk smith, trigger) spreads through the building adjacency graph, biased by density, dryness (drought!), and wind direction; rain suppresses. Burned lots → charred ruins → rebuild timer gated on prosperity.
Flood: winter snowpack meter; heavy melt + spring storms swell the river 1.5× width, damaging riverside lots and bridges.
Drought: summer event; harvest ×0.4, grass palette to straw, fire dryness up → feeds the hunger cascade.

Earthquake: rare; camera shake, 2–8% of buildings near epicenter to rubble, secondary fires.

Bandits: camps spawn in forest far from patrol coverage when unrest is high or war rages; visible caravan ambushes (train halts, scuffle, goods lost); crown posts bounties and dispatches patrols → skirmish → camp cleared or a Bandit King notable emerges if ignored.

The dragon (Myth ≥ Low): lairs in the highest peak at worldgen. Wake probability scales with treasury size. A raid is a Director-mandatory beat: wide glide in, torches a district (reuses fire system), takes gold, returns. The crown responds per monarch valor — muster and attempt a slaying (victory = hoard windfall + "Dragonslayer" epithet + legitimacy surge; failure = a dead king) or pay tribute (treasury drain, 5 years' peace).

Omens: comet (visibly crosses the sky), eclipse (real-time lighting event), red moon, two-headed calf. At Myth Low they move unrest and flavor the Chronicle; at Myth High they have mechanical teeth (a comet year multiplies event rolls) and a wandering prophet notable may appear.
Myth dial (Off / Low / High, default Low) lets the same engine serve grounded-medieval and high-fantasy moods.
12. The Chronicle
A scrolling parchment panel writing dated entries in annalist voice via a template grammar with slots, tone variants, and occasional editorializing ("…and some say the comet foretold it"). Entry classes: births/deaths/marriages/coronations, harvests and famines, trade and founding, war declarations/battles/sieges/peaces, disasters, omens, bandit and dragon affairs, festivals.
Every entry is clickable → camera flies to the location (if it still stands; else to the ruin or cairn).

Reign headers divide the scroll: "Here begin the years of Queen Maelis, second of her name."

Filter chips (Crown / War / Trade / Fates). Export button saves the full annals as .txt — a shareable artifact of the run.

Core is fully local (no API). Optional enhancement when running as a Claude artifact: a "Illuminate this entry" button that expands one entry into a paragraph via the API — omitted from the Netlify build.

13. The auto-director
The feature that makes it fascinating to watch. An event bus emits beats with priorities (dragon raid 10, war declared 10, coronation 9, battle 8, great fire 8, plague arrival 7, royal wedding 6, founding 6, caravan ambush 5, festival 3). The director:

Holds shots 8–20 s with eased fly-to transitions; per-category cooldowns prevent fire-fixation; higher priority interrupts lower.

Uses shot grammar: wide establishing → push-in on the subject; low tracking shots for marches and caravans; slow orbit for sieges and festivals; high wide for territory flips.

When idle, drifts between golden-hour establishing orbits of settlements.

Renders a lower-third beat card with the Chronicle line for the current shot.
Watch mode (C): hides all panels except the date and beat card. This is the screensaver-documentary mode — the whole app collapses to pure spectacle. Any manual camera input politely suspends the director for 30 s.
14. Control panel
Left drawer, four tabs.

Time (always visible in HUD): ⏸ ▶ ▶▶ ▶▶▶ ▶▶▶▶ mapping to 0 / 0.5 / 2 / 8 / 30 sim-days per second.

Rates — sliders, ×0.25–×4.0 multipliers, default ×1.0 unless noted:

Slider Governs Harvest yield food production Birth rate / Mortality population pool dynamics Trade volume caravan/cog spawn rate Tax rate (5–30%, default 12%) treasury vs. loyalty & unrest Plague virulence / lethality β and death fraction Disaster frequency lightning, quake, flood, drought rolls Bandit activity camp spawn & raid rate House aggression loyalty drift & war appetite Royal fertility dynasty robustness (low = succession drama) Myth Off / Low / High
Acts — trigger buttons (⌖ enters cursor-placement mode): Unleash plague ⌖ · Start a great fire ⌖ · Flood the river · Declare a drought · Earthquake ⌖ · Wake the dragon · Raise a bandit camp ⌖ · Assassinate the monarch · Contest the succession · Arrange a royal marriage · Bless the harvest · Send a comet · Found a village ⌖ · Grant 5,000 gold · Raze building ⌖.
Overlays (one at a time, ground-projected tint + line layers): Territories (house colors, borders animate during war) · Trade (flow lines, width = volume) · Prosperity heat · Plague spread · Unrest · Terrain.
World: seed field + Reforge · Copy share link · Export chronicle · Toggle labels · Watch mode.
15. HUD & inspector

Top strip: date & season glyph · weather icon · monarch name + sigil chip · treasury · realm population · time controls. Right: the Chronicle scroll. Left: control drawer (collapsible). Everything hides with H.

Inspector card (raycast click): buildings (name, type, tier, household count, stored goods); settlements (population, prosperity, unrest, food stores, walls, garrison, banner); notables (sigil, age, traits, spouse/heirs, grudges, current agenda); armies (strength, morale, supply, commander); caravans (route, cargo, value); the dragon (name it — dragons deserve names — mood, hoard). A compact Houses panel shows the five sigils, loyalty bars, and the line of succession — enough dynastic legibility without a full family-tree UI.

16. Data model sketch
Plain objects, no classes-for-their-own-sake:
World { seed, heightmap, biomes, rivers[], settlements[], roads[], houses[], notables[], armies[], caravans[], threats[], clock, treasury, legitimacy }
Settlement { name, pos, pop, prosperity, unrest, stores{7}, walls, district[], buildings[], owner(house), infected{S,I,R} }

Building { archetype, tier, lot, state(sound|burning|ruin|rubble), emissive }

House { name, sigil, seatSettlement, loyalty, might, wealth, grudges{} }

Notable { name, house, age, sex, traits[2], spouse, heirs[], role, agenda, alive }

Army { house, strength, morale, supply, commander, path, departT, arriveT }
Caravan/Cog { origin, dest, cargo, value, path, departT, arriveT }
Event { day, class, priority, pos, actors[], text } → feeds Chronicle + Director.
Systems are pure-ish functions over World called in tick order. One rand(stream) gateway for all stochastic draws.
17. Explicit non-goals (cut list)
No per-commoner simulation beyond bubble ambience. No naval combat (cogs are traders). No building interiors. No history replay or save-states (the seed regenerates year 0; the exported chronicle is the record). No diplomacy UI beyond marriages and loyalty. No frameworks, no server, no assets. Battles resolve abstractly under good visual dressing — this is a chronicle, not a wargame.
18. Build phases (each ships a runnable artifact)
Phase Scope Acceptance 1 — The Land Gen pipeline, terrain/water/sky, roads & bridges, buildings + full LOD, camera rig, day/night, seasons, seed URL Fly from 3 km to a market stall at 60 fps; two seeds feel like different countries 2 — The Wheel Clock & speeds, economy, caravans & cogs, population pools, settlement growth/decay, HUD, inspector, Chronicle v1, Rates panel Left at ▶▶▶ for 10 min: visible growth, ≥15 chronicle entries, no runaway values 3 — The Crown Notables & dynasty, houses & heraldry, loyalty/legitimacy/unrest, succession, civil war, armies/battles/sieges, territory overlay, political Acts "Assassinate the monarch" with rival heirs → war within a sim-year, visible marches, a peace, and a coherent chronicle arc 4 — The Fates Disaster suite, bandits, dragon & Myth dial, omens, auto-director + Watch mode, remaining Acts/overlays, ambient audio (optional), perf pass The five-minute test

19. Acceptance: the five-minute test

Defaults, speed ▶▶▶, no input. Within 5 minutes a viewer sees: a season change; ≥8 chronicle entries including ≥1 named-character event; ≥1 visible incident (fire, ambush, omen, festival) that the director covers; and at least one district visibly grow. Within 30 minutes: a succession or a war. After 200 unattended sim-years: population, treasury, and house count remain in sane bounds and the realm is still producing story.

If the five-minute test passes, the thing is fascinating to watch. Everything above serves that sentence.

Claude Fable 5 access just got extended until July 12.

5 more days to escape the permanent underclass.

Ramp 🤝 Replit

Powered by Ramp, business builders on Replit can now spin up incorporation, banking, cards, and bills straight from Replit Agent.

I heard it didn't work great on phones, I told Fable. Now it does.

Also signs and portents, royal processions, mule trains, bandit camps, tiny sheep, rivers, plagues, assassinations, marriages, fields, natural resources, and other stuff.

Meta's newest AI release: Muse Image

Powering creative tools in Meta AI and integrated into areas like Instagram Stories and WhatsApp chats.

No benchmarks published yet

I released sqlite-utils 4.0, the 124th release but the first major version bump since 3.0 back in 2020

I managed to keep things backwards-compatible all the way up to version 3.39 before the accumulated design mistakes forced me to bump that number!
simonwillison.net/2026/Jul/7…

Upgrade to 50,000 daily, thank you
@Cloudflare
!!!

Wan-Streamer v0.2

Higher Resolution, Same Latency

I built this beautiful AC automation with Claude Code + Home Assistant, perfect temperature all day based on temperature and humidity sensors, and then the sun position hitting at 5pm low into the room, etc. all good but yeah it sends a lot of requests to keep temp stable
Everything I do is on a VPS these days, I do nothing local anymore

Local AI coding just costs me battery and slower deploy time vs coding on the VPS with no battery spend and instant deployment!

Why deploy when you don't have to deploy!

Content Seal is built into the Meta AI app and
meta.ai
— every generated image carries a hidden provenance signal that stays intact through cropping, compression, and resizing.

We are previewing a public identification tool to verify the presence of this watermark on images created by the Meta AI app or
meta.ai
:
meta.ai/identification

Mac trackpads fixed.

I should just have Fable monitor the thread for bug reports and solve them.

PixWorld

Unifying 3D Scene Generation and Reconstruction in Pixel Space

Here was the original post (there were a few others as well)!

Impressive
@arena
benchmarks for Muse Image, slotting in below just OpenAI's GPT Image 2 on text-to-image, single image editing, and multi-image editing.

Meta also previewed Muse Video, which debuts at an impressive No. 3 spot on the Video leaderboard.

🐙 Fall into the Backrooms with one prompt.

"Backrooms Dreamcore" is now live on MiniMax Hub's Skill Square.
Try and create your own dreamcore space!


#MiniMax
#Hailuo
#MiniMaxHub

The goal here is to radically reduce the cost, friction, and complexity of putting extremely capable agents into production.

We have already seen thousands of customers pushing the frontier with this API.

The roadmap for managed agents is super ambitious, with a new UI experience, new customer agents built by Google, and so much more coming soon! Great work by
@alihcevik
@_philschmid
@mcocirio
and many others to make this happen.

Here is the code, if you want to play with it:
github.com/emollick/annals-k…

Worked!
A more European reply does not exist
2 mg Melatonin is 10x too much

Melatonin is oddly sold in 10-100x too high doses

The right dose for Melatonin is about 0.2-0.3mg or 200-300mcg

Very very low taken an hour before sleep to give your natural melatonin production a kickstart, not replace it!

Gemma 4 Technical Report

New this week 🚀 Replit Community Profiles — Proof of Work for vibe coders.

Your profile, your flex. Get an activity graph of your agent usage and checkpoints, plus a Replit Power Ranking for pro users.

Log in, claim your profile, pick your best projects, and share your stats with friends.

See it in action now →
replit.com/community

Now I'm regretting having carefully blasted through 100% of my Fable allotment in time for being cut off at the end of today!

(Presumably "Wed 12:00 PM" means I get it back at noon on Wednesday, not at midnight between Tuesday and Wednesday?)

Proof of Work for Professional Vibe Coders
nitter.tiekoetter.com/i/broadcasts/1RKZzzLbQ…

Muse Image demonstrates emergent self-refinement within its chain of thought, adaptively executing local edits, complete re-generation, or tool use. Rather than being explicitly programmed, this behavior emerged during RL training as a strategy to optimize image quality and maximize reward.

It runs on wheeled, legged, and flying robots and generalizes across sizes, unlocking delivery, logistics, manufacturing, and hospitality.
Read more:
mistral.ai/news/robostral-na…
Trained entirely in simulation: ~400,000 trajectories across 6,000 scenes. A prefix-caching recipe cuts training tokens by 22×, turning months-long runs into days. Online RL (CISPO) pushes success rates higher still.
76.6% success on R2R-CE validation unseen (79.4% on validation seen), the benchmark for following instructions in previously unseen environments. It beats the best single-camera approach by 9.7 points while using far less sensing.
No LiDAR. No depth sensors. No camera rig. Where leading systems lean on depth or multiple cameras, Robostral Navigate works from one ordinary RGB camera, and still comes out ahead.

Since this is a slightly backwards-incompatible change here's a detailed upgrade guide - which you can read, or feed into your coding agent and have it apply the upgrades for you
sqlite-utils.datasette.io/en…

DoorDash just published new AI research.

Yes, the delivery app. Its AI research team built an internal test: 105 past code changes from its own engineers, re-run through AI code reviewers to see how many real problems each one catches.

The reviewer works as a pair: one model skims every code change and flags suspicious spots, a second digs into each flag. Today, both jobs go to Anthropic models.

But that test found a cheaper split: hand the skimming to Kimi K2.6, an open Chinese model, and save Claude Fable 5 for the hard part.

The K2.6 / Fable split scored 65.2% at catching real problems (worse bugs count for more) vs. 53.6% for the all-Anthropic pair. It caught 8/10 critical bugs, and cost $3.81 per code change instead of $3.91.

Co-founder Andy Fang says the test, called DashBench, is what lets them bring in open models.

Fang: "Better quality, cheaper cost."

Model page:

ollama.com/library/glm-5.2

🏆 Kling AI NEXTGEN Awards Ceremony | Winners Unveiled

From campus creativity to cinematic 4K productions, this year's winning works reflect the breadth of AI filmmaking today. Through their films, creators explore real-world issues, personal emotions, and new possibilities for visual storytelling—using AI as a tool to turn creative ideas into powerful stories.

Congratulations to all the winners! Thank you to every creator for sharing your vision—we can't wait to see what's next.

Swipe through to explore the full winner list.

Another powerful new AI image model has arrived!

ByteDance released Seedream 5.0 Pro, a model that claims to go beyond generation to "understand design".

Features include:

- Improvements in text rendering, structure, alignment, and other design areas for strong infographics and prototypes

- Precision editing with layer seperation for editable designs, and the ability to combine multiple images

- Upgrades to realism, lighting, skin textures

- Multilingual inputs and generations

I had Claude Fable 5 run a final review of the release using this prompt and it kicked up another four release blockers!
simonwillison.net/2026/Jul/7…

Seedance 2.0 Mini is now available in Luma.

Generate video fast enough to keep up with how you actually work. Storyboard a concept, test a dozen directions, and land on the cut before the idea cools.

Powerful output. Lighter footprint. Ready for your workflow right now.

Open Luma and put it to work →
lumalabs.ai/app
.

Top stories in AI today:

- Meta’s new in-house AI image model
- Beijing considers new restrictions for Chinese AI
- Run better 1-on-1s with employees using AI
- DoorDash grades the AI reviewing its code
- 4 new AI tools, community workflows, and more

Muse Image learns to search the web to ground generated images in factual and real-time information and visual references. Enabling search improves the accuracy of generated images for categories related to real-world knowledge and factuality.

Explore this thread:
meta.ai/share/c/ZTPQAeJh4e

paper:
huggingface.co/papers/2607.0…

Of course this isn’t activation access. It’s structured self-report, the paper’s own finding is that what Claude reports is causally read out of its workspace. Nonetheless, it is a bit of light into models own process.

You can download the skill here.

github.com/Doriandarko/skira…

paper:
huggingface.co/papers/2607.0…

paper:
huggingface.co/papers/2607.0…